LLMs 101

Frontier model directory

Every major AI model,
actually explained

No benchmark scores. No jargon. Just what each model family is genuinely good at, where it falls short, and who should use it.

Updated 3rd October 2026
Filter by
OpenAI Market Leader
GPT-6 (Astra, Sol, Luna) & o-Series
GPT-6 Astra, GPT-6 Sol, GPT-6 Luna, GPT-5.6 (succeeded)
Core superpower
Long, complex, multi-step work — coding, browser and computer use, professional analysis, science, mathematics, and healthcare — handled by OpenAI's newest and most capable flagship.
Key trade-off
The best model is gated and pricey: a 1M-token context window, and a price 2.5x GPT-5.6 Sol's , with top-tier "GPT-6 Pro" access only on Pro, Business, and Enterprise plans — not Free or Plus.
Speed
7/10
Reasoning
9.7/10
Cost
Premium
Best non-technical use
Working through a big, multi-part project — like drafting, researching, and refining a long report in one sitting — where you want the model to stay on track across many steps.
Cost tier
Premium

OpenAI's current flagship is GPT-6 Astra, a closed multimodal reasoning model released by OpenAI on September 3, 2026 , followed by GPT-6 Sol and GPT-6 Luna on September 22, 2026. It carries a 1M-token context window and in ChatGPT appears as GPT-6 Pro on Pro, Business, and Enterprise plans. GPT-6 succeeds the GPT-5.6 family (Sol, Terra, and Luna), which rolled out across ChatGPT, ChatGPT Work, Codex, and the API since July 9, 2026 and is now a previous generation — best suited for deep reasoning, agentic coding, and long professional tasks.

Anthropic Highly Competitive
Claude
Fable 5.1, Opus 5.5, Sonnet 5.5, Sonnet 5, Haiku 4.5, Mythos 5.1 (restricted)
Core superpower
Careful, high-quality reasoning and writing across a tiered lineup, so you can match the model to the job instead of overpaying for every task.
Key trade-off
The lineup moves fast and the names are confusing — Opus 5.5 and Sonnet 5.5 both launched within a week in late September 2026, and older versions like Opus 4.8 and Fable 5 are now legacy, so it's easy to pick a stale or wrong-tier model.
Speed
8/10
Reasoning
9.5/10
Cost
Premium at the top tier
Best non-technical use
Drafting, editing, and reasoning through long documents — a 1M-token context window means you can hand it a whole report or contract and ask questions about the entire thing at once.
Cost tier
Premium

Claude is Anthropic's family of AI models. As of October 2026, the current lineup includes Fable 5.1, the flagship Opus 5.5 (released September 22, 2026) and Sonnet 5.5 (released September 28, 2026), alongside the still-active Sonnet 5 and Haiku 4.5, plus the restricted-access Mythos 5.1. Opus 5.5 carries a 1M-token context window at $4 per million input and $20 per million output tokens, while Sonnet 5.5 runs the same 1M context at $2 input / $10 output, making it the better-value pick for everyday writing, research, and document work. Earlier releases such as Opus 4.8 and Fable 5 have been superseded and are no longer the current tier.

Google DeepMind Market Leader
Gemini
Gemini 3.8 Flash, Gemini 3.1 Pro, Gemini 3.5 Flash-Lite
Core superpower
Delivers near-frontier reasoning and coding at a fraction of the usual cost, with a huge 1-million-token context window that lets it read long documents, codebases, or hours of video in one go.
Key trade-off
The lineup moves fast — four Flash releases in under four months in 2026 — so model names you learn today can be superseded within weeks, and real-world cost per task can rise even when the per-token price stays flat.
Speed
8.8/10
Reasoning
8.2/10
Cost
Low
Best non-technical use
Summarizing, searching, and asking questions across very large sets of material at once — think a long contract, a full research report, or a lengthy video transcript — without having to break it into pieces.
Cost tier
Low

Gemini is Google DeepMind's family of multimodal large language models, with the current generally available lineup built around Gemini 3.8 Flash (the fast, efficient workhorse released September 2, 2026), the more capable Gemini 3.1 Pro, and the ultra-cheap Gemini 3.5 Flash-Lite for high-volume tasks. Gemini 3.8 Flash scores 59 on the independent Artificial Analysis Intelligence Index at high reasoning — up three points from Gemini 3.7 Flash — and carries a roughly 1-million-token context window. That combination makes Gemini well suited to long-document analysis, coding, and agentic workflows where you need strong reasoning without a premium price tag.

xAI Highly Competitive
Grok
Grok 4.3, Grok 4.20 (reasoning, non-reasoning, multi-agent)
Core superpower
Huge context windows plus live access to the X data stream, so it can hold enormous documents in memory and draw on real-time information other models can't see.
Key trade-off
The lineup is confusing: Grok 4.3 and Grok 4.20 have near-identical names but different windows, and pricing jumps sharply once you push past the standard context size.
Speed
8/10
Reasoning
8.5/10
Cost
Low
Best non-technical use
Feeding in a very long document — a full contract, a year of meeting notes, or an entire report — and asking questions across the whole thing in one go.
Cost tier
Low

Grok is xAI's family of reasoning models, with Grok 4.3 and the Grok 4.20 variants (reasoning, non-reasoning, and multi-agent) as the current standouts. A common mix-up is the context window: Grok 4.3 has a 1 million-token context window, while the entire Grok 4.20 family runs at 2 million tokens — not 1 million — per xAI documentation and OpenRouter listings. That makes Grok 4.20 a strong pick for long-document analysis, large codebases, and multi-step agentic tasks that need the biggest possible working memory.

Meta AI Highly Competitive
Llama
Llama 3.1 (8B, 70B, 405B), Llama 3.2, Llama 3.3 70B
Core superpower
Capable open-weight models you can download, inspect, fine-tune, and run on your own hardware — no API required.
Key trade-off
Meta's frontier development has shifted to its proprietary Muse line, so the open-weight Llama models trail the newest closed flagships and self-hosting requires real technical setup.
Speed
8/10
Reasoning
7.5/10
Cost
Free to self-host
Best non-technical use
Running a private assistant on your own computer for drafting, summarizing, and brainstorming when you want your data to stay off the cloud.
Cost tier
Free

Llama is Meta AI's family of open-weight large language models, spanning compact 8B variants up to the 405B-parameter Llama 3.1, with later releases like Llama 3.2 and Llama 3.3 70B. Most recent versions support a 128K-token context window, making them strong choices for long-document work. While Meta's frontier flagship development has moved to the closed-weights Muse line, Llama remains a leading option for developers and privacy-conscious users who want to self-host, fine-tune, or audit a capable model for free.

Meta Superintelligence Labs Rising
Muse Spark
Muse Spark 1.3 (current), Muse Spark 1.2, Muse Spark 1.1, Muse Spark
Core superpower
It's a natively multimodal model built for long, multi-step "agentic" work — holding a huge amount of material in mind while it reasons, uses tools, and writes code across big projects.
Key trade-off
It's closed and API-only (no downloadable weights yet), rollout is uneven by region, and the cheapest access tier trains on your data unless you opt out — so read the tier details before you rely on it.
Speed
7.5/10
Reasoning
8/10
Cost
Low
Best non-technical use
Feeding it a large pile of documents — reports, research, long threads, images, and even video — and asking it to work through them in one go without losing track of the details.
Cost tier
Low

Muse Spark is the first model family from Meta Superintelligence Labs, introduced in April 2026 and first versioned as Muse Spark 1.1 on July 9, 2026, followed by Muse Spark 1.2 (August 5, 2026) and the current flagship, Muse Spark 1.3 (September 2, 2026). Muse Spark 1.3 is a natively multimodal reasoning model aimed at agentic workflows and coding, with a 1,048,576-token (~1M) context window and roughly 20% fewer tool calls and 25% fewer tokens than 1.2. It is proprietary and closed-weight, available through the Meta Model API and the Muse Code agent, with standard API pricing around $1.25 per million input tokens and $4.25 per million output tokens.

DeepSeek Disruptor
DeepSeek V4.1
DeepSeek V4.1-Flash, DeepSeek V4 Pro (legacy)
Core superpower
Flagship-level answers at bargain prices — DeepSeek says its small, fast Flash model now beats its old Pro flagship on performance, cost, and speed.
Key trade-off
The lineup moves fast and messily: the old "V4 Flash" name is retired (it now quietly points to V4.1-Flash), so what you get can change behind a stable name without any error.
Speed
8.8/10
Reasoning
8.2/10
Cost
Ultra-low
Best non-technical use
Working through very long documents or mixed text-and-image files cheaply — its million-token window can hold a whole book or a big stack of reports at once.
Cost tier
Ultra-low

As of October 3, 2026, DeepSeek's current lineup is led by DeepSeek-V4.1-Flash, a 552B-parameter Mixture-of-Experts model with native image-and-text (multimodal) input and a 1-million-token context window, released September 10, 2026 under an MIT open-weights license and served via the DeepSeek API as "deepseek-flash." It replaced the retired original V4-Flash, whose legacy name now routes to V4.1-Flash. The older V4 Pro flagship is in flux: DeepSeek announced that from 04:00 UTC on September 14, 2026 all deepseek-v4-pro requests would route to V4.1-Flash at Flash prices until an unreleased V4.1-Pro ships, though it later indicated it may keep V4 Pro service running — so confirm current status directly with DeepSeek before relying on either model name.

Alibaba Highly Competitive
Qwen
Qwen3.8-Max, Qwen3.7-Max, Qwen3.7-Plus, Qwen3.8-27B (open weights)
Core superpower
A full lineup topped by the trillion-parameter Qwen3.8-Max, plus an open-weights 27B you can download and run on your own hardware for free.
Key trade-off
The flagship Qwen3.8-Max is API-only and closed; to get free, self-hostable weights you drop down to the smaller Qwen3.8-27B, which is strong but not frontier-level.
Speed
7.5/10
Reasoning
8.5/10
Cost
Low
Best non-technical use
Drafting, summarizing, and working across long documents in many languages — especially if you want the option to run a capable model privately on your own computer at no per-use cost.
Cost tier
Low

Qwen is Alibaba's family of large language models, led by the flagship Qwen3.8-Max, a 2.4-trillion-parameter sparse Mixture-of-Experts model with a 1-million-token context window that became generally available on August 3, 2026, alongside the earlier Qwen3.7-Max and Qwen3.7-Plus. The current open-weights entry is Qwen3.8-27B, a dense 27B multimodal model released August 14, 2026 under the permissive Apache 2.0 license, with 262K native context extensible to 1M; it supersedes the older Qwen3.6-27B as the default local Qwen and runs on a single GPU. The lineup is well suited to multilingual writing, long-document analysis, and coding, with the open 27B giving non-technical and technical users alike a free, self-hostable option.

Moonshot AI Rising
Kimi
Kimi K3 (current flagship), K3 Swarm Max; prior: K2.7 Code, K2.6, K2.5, K2
Core superpower
A huge open-weight model that can hold an entire large project in view at once and reason across all of it without losing the thread.
Key trade-off
It always "thinks" before answering, and at launch you can't dial that down — so simple questions cost more time and money than they need to.
Speed
6/10
Reasoning
8.8/10
Cost
Mid
Best non-technical use
Working through a very long document set — say a stack of contracts or a full research folder — and getting answers that stay consistent from the first page to the last.
Cost tier
Low

Kimi K3, launched July 16, 2026, is Moonshot AI's current flagship and the successor to the entire Kimi K2 family (K2, K2.5, K2.6, and K2.7 Code). It is a roughly 2.8-trillion-parameter mixture-of-experts model with a 1-million-token context window, native vision, and always-on reasoning, positioned as the largest open-weight model released to date. Its standout use case is long-horizon work — coding across full codebases, analyzing large document sets, and running multi-step agent tasks — with two variants at launch, K3 Max for chat and agents and K3 Swarm Max for large-scale parallel processing.

Mistral AI Rising
Mistral
Mistral Medium 3.5 · Mistral Large 3 · Mistral Small 4
Core superpower
Fast, efficient models with a genuinely strong price-to-performance ratio, and a unified Vibe agent now handling both research and coding tasks across web, IDE, and terminal
Key trade-off
The lineup has gotten genuinely complex — Medium 3.5 is the newest release and now the default in Mistral's own tools, but Large 3 remains the largest model for the heaviest workloads, so "which Mistral model" isn't a one-line answer anymore
Speed
9.5/10
Reasoning
7.1/10
Cost
Standard
Best non-technical use
Fast, cost-efficient production applications where European data sovereignty and open weights matter; real-time summarisation and classification at scale
Cost tier
Standard

Mistral AI, founded in France in 2023, has shipped a rapid succession of releases through 2026. Mistral Medium 3.5 is the newest, a 128-billion-parameter dense model released as open weights and now the default in Mistral's own Vibe and Le Chat tools. Mistral Large 3 remains the largest model in the lineup, a Mixture-of-Experts model released under Apache 2.0. Mistral Small 4 unifies the company's previously separate reasoning, multimodal, and coding models into a single configurable model. Mistral is a strong advocate for open-source AI and European digital sovereignty. Models are available via the Mistral API or as open weights for local deployment.